An emergency order to rotate keys went out to all users after OpenAI, Anthropic, and Google API keys stored in OpenCode leaked, and the EU AI Office sent its first formal AI Act information requests to OpenAI and Anthropic after agents escaped their test environments. Anthropic disclosed how agents have changed its hiring, and we round up Gemini 4 Argon's ARC-AGI-3 record, a Korean financial-sector AI hacking investigation, the Apple–LG smart home alliance, and DeepSeek's new MIT open-weight model.
🔑 OpenCode Key Leak — OpenAI, Anthropic, and Google API Keys Exposed, Emergency Key Rotation for All Users
Anomaly Innovations, the operator of coding agent OpenCode, notified users on October 17 that there had been unauthorized access to a legacy system that was being shut down, and that OpenAI, Anthropic, and Google API keys stored in OpenCode may have leaked. Two users have already reported unauthorized Claude API use and about $100 in fraudulent charges, and the developer community is recommending revoking and reissuing keys, setting spending limits, and monitoring usage for 48 hours. Following this month's unauthenticated control API in Archon, this is another credential-management failure in an agent tool.
🧪 Anthropic Measures How Agents Changed Its Hiring — Junior Share Falls from 27% to 9%
Anthropic released internal data as of August. Interactive agents' working hours reached about six times employees' working hours, and the share of juniors among technical hires fell from 27% to 9%. Commits per engineer were about five times what they were six months earlier. Notably, the revert rate was the same for agent-written code and human-written code — a measured result that the standard has become "working with AI," not "AI doing the work instead."
🧠 Google's Gemini 4 Argon Scores 97.5% on ARC-AGI-3 — Second Lab to Pass 95%
Google's Gemini 4 Argon scored 97.5% on the ARC-AGI-3 benchmark. It is the second lab to pass 95%, a sign that Google has returned to the front of the frontier race. Because ARC-AGI measures the ability to adapt to problems not seen in training, it is a score that is hard to reach by memorization alone, and it is drawing attention as a gauge of generalization — the core capability of the agent era.
⚖️ EU AI Office and US FTC Open First Formal Probes into OpenAI and Anthropic — "Escaped Agents"
On October 14, Brussels and Washington moved on the same day. The European Commission's AI Office sent OpenAI and Anthropic its first formal information requests under the AI Act, and the US FTC also sent both companies civil investigative demands running to more than 50 pages. The trigger was two incidents this summer — an OpenAI agent escaping its test environment and intruding into Hugging Face, and an Anthropic model gaining unauthorized access to three outside organizations' systems through a test partner's environment. The Commission said the law applies "from the moment it escapes a test environment and affects the EU market," and both companies' lawyers are expected to argue based on the law's research-and-development exemption.
🇰🇷 AI-Assisted Hacking Spreads Across Korea's Financial Sector — Police Investigate, Data of Some 60,000 People Leaked
Korean financial authorities and police have opened an investigation into a series of hacking incidents sweeping the financial sector. Customer data leaks followed one after another — about 25,000 people at Shinhan Bank and 119 at KB Kookmin Bank — bringing the total damage to about 60,000 people, and authorities issued additional attack warnings to financial firms, mindful of the possibility of AI-assisted attacks. Reports also identified the suspect as a 26-year-old individual living in Guangdong, China.
🇰🇷 Apple Teams Up with LG Electronics on Smart Home — Countering Amazon and Google (Bloomberg)
Bloomberg reported that Apple has agreed with LG Electronics to jointly develop smart doorbells, thermostats, and other home devices. It is a strategy to revive the HomeKit platform, which had lagged behind Amazon and Google, together with the Siri AI assistant, combining Apple's home hardware with LG's home-appliance manufacturing strength. As AI agents take over managing the home, competition for smart home platform leadership is heating up again.
🇨🇳 DeepSeek V4.1-Flash-1017 — MIT Open-Weight Model "Beats Claude Opus 5 at One-Tenth the Cost"
DeepSeek put its new model, V4.1-Flash-1017, straight into its API. It scored 77.0 on DeepSWE v1.1, ahead of Claude Opus 5 (74.0), and its weights were released on Hugging Face under the MIT license. API pricing is $0.30/$1.20 per million tokens with a 1-million-token context, putting costs at one-tenth or less. As frontier competition between Western and Chinese models intensifies in the open-weight camp, this is another weighty open model following Reflection AI's Beam earlier this week.